AI は日立の注力・差別化技術であり、世界全体のデジタルトランスフォーメーションをリードする可能性が期待されています。日立のあらゆる事業において、事業の成長拡大に貢献するためには、強力で有用な AI 技術が必要です。グローバルで多様なチームの中で、先見性を持って社会の解くべき課題を見極め、先端AI技術を駆使して解決します。最新の市場や研究成果の動向をとらえ、顧客目線で顧客事業の成長に貢献するように、自社製品やサービスの拡張や革新を提案して、AIアルゴリズムの開発からデモンストレーションまでの必要な研究開発を行います。データから現実世界で起こっていることを分析し、予測に基づいて最適化の方策を導き出し、制御することで全体としての改善に繋げていきます。その様々なシーンで活躍するAI技術を開発します。
機械学習、Deep Learning応用、IoT、地理空間情報、データ解析、人・モノの流れ分析、モデリング・シミュレータ、デジタルツイン、機械運転自律化、プロセス最適化
Osakabe, Y.; Asahara, A. MatVAE: Independently Trained Nested Variational Autoencoder for Generating Chemical Structural Formula. AAAI Spring Symposium: MLPS 2021
http://ceur-ws.org/Vol-2964/article_69.pdf
Abstract: Rapid materials development utilizes deep generative models to suggest candidate compounds with desirable properties before actual experiments. Such models successfully generate novel candidates with improved properties in some cases, but they usually require a large experimental dataset which is difficult to obtain. We propose MatVAE–two nested VAEs independently trained on different datasets. The first VAE, which is trained on a huge open dataset, is a universal generator of chemical structural formulae, and the second VAE, which is trained on a small experimental dataset, learns the structure–property relation. This training framework can be understood as a semi-supervised learning, which is expected to enhance model transferability. We verified that MatVAE generates five times more valid candidate compounds than the conventional un-nested single VAE model.
Khorasgani, H.;Wang, H.; Gupta, C.; Serita, S. Deep Reinforcement Learning with Adjustments. 2021 IEEE 19th International Conference on Industrial Informatics (INDIN), 2021
https://ieeexplore.ieee.org/abstract/document/9557543
Abstract: Deep reinforcement learning (RL) algorithms can learn complex policies to optimize agent operation over time. RL algorithms have shown promising results in solving complicated problems in recent years. However, their application on real-world physical systems remains limited. Despite the advancements in RL algorithms, the industries often prefer traditional control strategies. Traditional methods are simple, computationally efficient and easy to adjust. In this paper, we first propose a new Q-learning algorithm for continuous action space, which can bridge the control and RL algorithms and bring us the best of both worlds. Our method can learn complex policies to achieve long-term goals and at the same time it can be easily adjusted to address short-term requirements without retraining. Next, we present an approximation of our algorithm which can be applied to address short-term requirements of any pre-trained RL algorithm. The case studies demonstrate that both our proposed method as well as its practical approximation can achieve short-term and long-term goals without complex reward functions.
I. Suemitsu, H. K. Bhamgara, K. Utsugi, J. Hashizume and K. Ito, "Fast Simulation-Based Order Sequence Optimization Assisted by Pre-Trained Bayesian Recurrent Neural Network," in IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7818-7825, July 2022, doi: 10.1109/LRA.2022.3185778.
https://ieeexplore.ieee.org/document/9804857
Abstract: This paper presents a fast optimization method for the picking order sequence of automated order picking systems in logistics warehouses. In this order sequencing problem (OSP), the fulfillment sequence of the given picking order set is determined to optimize the performance measures such as makespan and deadlock occurrence. Simulation is generally necessary to evaluate these measures for complex automated systems. However, their order sequence cannot be optimized quickly due to the long calculation time. It may make the system productivity and flexibility lower than expected because its picking schedules cannot be updated frequently. We, therefore, propose a fast optimization method to solve these simulation-based OSPs by taking a pretrained surrogate-assisted optimization approach. Firstly, we utilized a Bayesian recurrent neural network (BRNN) as a surrogate model to accurately learn the relationship between picking order sequence and performances. Secondly, we developed the surrogate-assisted optimization method based on simulated annealing (SA) and BRNN. Numerical experiments show that the surrogate model can evaluate about 10000 times faster than the simulation. The proposed method also obtains an optimized solution 8.9 times faster than simulation-based optimization by the original SA.
Kato, T.; Kamoshida, R. Multi-Agent Simulation Environment for Logistics Warehouse Design Based on Self-Contained Agents. Appl. Sci. 2020
https://doi.org/10.3390/app10217552
Abstract: It is generally difficult to analyze the performance of a multi-agent system,thus it is important to model a warehouse and conduct simulations to design and evaluate the possible system configurations. However, the cost of modelling warehouses and modifying the models is high because there are various components and interactions compared to conventional multi-agent simulations. We proposed a self-contained agent architecture and message architecture of a multi-agent simulation environment for logistics warehouses to reduce the simulation-model development and modification costs.
Yawata, K., Osakabe, Y., Okuyama, T., & Asahara, A. QUBO-inspired Molecular Fingerprint for Chemical Property Prediction. In 2022 IEEE International Conference on Big Data (Big Data) (pp. 2437-2440). IEEE.
https://doi.org/10.1109/BigData55660.2022.10020236
Abstract: Molecular fingerprints are widely used for predicting chemical properties, and selecting appropriate fingerprints is important. We generate new fingerprints based on the assumption that a performance of prediction using a more effective fingerprint is better. We generate effective interaction fingerprints that are the product of multiple base fingerprints. It is difficult to evaluate all combinations of interaction fingerprints because of computational limitations. Against this problem, we transform a problem of searching more effective interaction fingerprints into a quadratic unconstrained binary optimization problem. In this study, we found effective interaction fingerprints using QM9 dataset.
“AI for Powering Good” – A joint workshop by Hitachi & Mila
https://www.hitachi.com/rd/sc/aiblog/011/index.html
Abstract: In October 2019, the “AI for Powering Good” workshop was organized as part of a research collaboration project which has been going on between Mila (Montreal Institute for Learning Algorithms) and Hitachi since spring 2019. The workshop was an idea to bring the researchers from both Hitachi and Mila to meet each other and to share and exchange knowledge as networking is a key element for successful collaboration.
地域密着型ショッピングモール施設の運営を効率化し、集客力アップに貢献するサイバーフィジカルシステム(CPS)を試作
https://www.hitachi.co.jp/rd/news/topics/2021/0614_nonowa_poc.html
概要: 日立は、地域活性化の起点となる地域密着型ショッピングモール向けに、施設運営のPDCAサイクルを支援するサイバーフィジカルシステム(CPS)のプロトタイプを開発しました。
Vegetation Manager: Manage tree growth before they cause power outages
https://www.hitachienergy.com/jp/ja/products-and-solutions/asset-and-work-management/lumada-fsm/vegetation-manager
Abstract: Utilities understand the importance of vegetation management to grid reliability and long-term resilience—but it’s an expensive, complex, and tedious process. Without the right data, you can’t plan; without planning, you can’t prioritize; without prioritizing, critical issues go unresolved, and a hazard tree takes down part of your grid. Be proactive with prevention.
疑似量子の活用で従来のMIによる材料開発期間をさらに20%短縮できることを実証
https://www.hitachi.co.jp/New/cnews/month/2022/12/1216.html
概要: 日立は、材料開発の加速につながる新たな機械学習モデルを開発し、積水化学工業株式会社と進めているマテリアルズ・インフォマティクス(以下、MI)推進に向けた協創活動においてその有効性を実証しました。本モデルは、決定木を用いた材料の性能予測モデルの構築に、量子コンピュータを疑似的に再現する「CMOSアニーリング」を適用することで、さまざまな条件を網羅的に考慮することが可能です。これにより従来の機械学習モデルの予測精度を本モデルの適用により向上させ、材料開発の期間を約20%短縮できる見通しを得ました。
